| 藤原 徹平 | 1751104: M, 1回目発表 | インタラクティブメディア設計学 | 加藤 博一 |
|
title: Pre-Attentive Contrast Cues By Computer Mediated Reality
abstract: An aging society is rapidly spreading in Japan and elderly people tends to have impaired physical function, especially visual function. Some related work shows our eye function can be affected by augmenting virtual cues in the view and the system can assist a part of eye function. Through these work, we think the system can improve user's basic eye vision and cognitive skills by subtly modifying what user can see. In order to prove this, we develop a display system that shows very small cues to user's view and investigate its effectiveness. language of the presentation: Japanese | |||
| 松村 遥 | 1751111: M, 1回目発表 | インタラクティブメディア設計学 | 加藤 博一 |
|
title: Can Face Swapping Technology Facilitate Mental Imagery Training?
abstract: Mental imagery training is usually done by athletes to improve their performance, which is sports training by observing an expert’s motion. As a basis for mental imaginary training, it is known that brain activity can be observed from a person watching a video of an athlete performing a motion, even if the said person is not performing the actual motion. In addition, observing one’s own video shows stronger brain activity than that of another person’s video. From these facts, we make the hypothesis that it becomes easy to imagine one's own motion by observing the another person’s motion video swapped the one's own face and face swapping facilitates athlete’s motor learning. In this presentation, I will show the result of the preliminary experiment to validate the effect of face swapping, and then discuss the future direction of this research. language of the presentation: Japanese 発表題目: 運動学習促進のための顔交換技術の検討 発表概要: スポーツトレーニングの現場において，しばしば熟練他者の動作を観察するイメージトレーニングによる運動学習が行われる．イメージトレーニングにより得られる運動学習効果の背景として，運動動作の映像観察時に動作の実施を想像することにより，（実際に動作は行っていないが）動作時に活動する脳の部位が活動することが知られている．さらに，他者動作映像観察時に比べ自己動作映像観察時の脳活動が大きいことが報告されている．そこ で，我々は，他者の運動動作映像に対してトレーニング対象者の顔を合成した映像を生成することによって，トレーニング対象者は容易に自分自身が動作を行っているイメージを持つことが可能となり運動学習をより促すのではないかと予想している．本発表では，初期検討として行った実験結果を示し，その考察および今後の展開について述べる． | |||
| 浅野 幸之助 | 1751002: M, 1回目発表 | 計算システムズ生物学 | 金谷 重彦 |
|
title: Staining Translation Model of Histpathological Image of Pancreas Cancer using Conditional Generative Adversarial Networks
abstract: In recent years, researches on image processing using deep neural network are rapidly proceeded. Along with that, the demand for Computer Aided Diagnosis(CAD) using AI is increasing on clinical site since there is a shortage of pathologists nationwide in Japan. My research focus on the phase of preparing histpathological specimen. I construct the model that predicts MT stained images from HE stained images using conditional generative adversarial networks. Also, I analyze what kind of features the network has learned using some clustering algorithms from outputs on hidden layers of the trained model. language of the presentation: Japanese | |||
| 井上 雄貴 | 1751011: M, 1回目発表 | 計算システムズ生物学 | 金谷 重彦 |
|
title: A prediction model of the effect of anticancer drugs based on the influence on metabolic pathway.
abstract: Cancer chemoprevention involves the chronic administration of a synthetic, natural or biological agent to reduce or delay the occurrence of malignancy. The potential value of this approach has been demonstrated with trials in breast, prostate and colon cancer. But the mechanism of how to kill cancer is not known yet. In this research, we are conducting research aiming at clarifying how ChemoPrevention kills cancer by using information science method. language of the presentation: Japanese | |||
| 本多 右京 | 1751107: M, 1回目発表 | 自然言語処理学 | 松本 裕治 |
|
title: Metaphor Identification Using Large Text Corpus
abstract: We are able to use metaphorical expression but mechines are not so far. One of the metaphor understanding task is metaphor identification, where a system judges whether a given phrase is metaphorical or literal. The difficulty of this task is its limited training data. Our proposal is to exploit large text corpus using knowledge base completetion method. language of the presentation: Japanese | |||
| 松野 智紀 | 1751110: M, 1回目発表 | 自然言語処理学 | 松本 裕治 |
|
title: Chunk Vector Expression for Neural Graph-Based Dependency Parsing
abstract: In recent years, Recurrent Neural Network based approaches have made significant success in graph-based dependency parsing. In spite of its considerable performance, all of the proposed methods use words as basic units for parsing, making it difficult to apply them to chunk based dependency parsing which is standard in Japanese parsing tasks. In this research, we propose novel method to obtain optimal chunk representation using method called “LSTM-Minus” and incorporating 2 types of Bi-LSTM layers, word LSTM and chunk LSTM. As a result, our model achieved comparable result with the state of the art Japanese syntactic parser JDepP. language of the presentation: Japanese | |||
| 和田 崇史 | 1751128: M, 1回目発表 | 自然言語処理学 | 松本 裕治 |
|
title: Unsupervised Sequence Matching
abstract: To extract relationship between data from different domains is a very important and challenging task, and a lot of work has tackled this problem. However, in order to find the relationship, most of the existing work requires a large amount of parallel data, which is very expensive and difficult to obtain. To solve this problem, we propose a method of matching sequential data from different domains without any parallel data. We define our model as “multi-domain sequence AutoEncoder”, a sequence AutoEncoder that is shared among data from multiple domains. It consists of LSTM Encoder-Decoder that is shared among all domains, and embeddings that are learnt independently of each other. The model architecture enables to map embeddings of different domains into a common space, and to learn common dynamics of sequences among all domains. In this work, we measure the effectiveness of our model by matching a sentence in one language and its translation in another. language of the presentation: Japanese 発表題目: 教師なし系列マッチング 発表概要: 異なるドメイン間のデータで対応関係を見つけることは, 長年取り組まれてきた重要な問題である. しかし, 多くの既存手法では人手で作成されたドメイン間でパラレルなデータを用いて対応関係を学習する ため, そのような教師データがない場合には適用できないという問題がある. そこで本稿では, パラレルな データを用いずに異なるドメインの系列データを対応づける手法を提案する. 全ドメインで共通の LSTM を持つ系列 AutoEncoder を学習することによって, 各ドメインの特徴量を共通の空間で表現する線形写像 を獲得し, 異なるドメインで共通した系列のダイナミクスを獲得することを可能にする. 複数言語の対訳文 のマッチングにより, 本提案法の有効性を示す. | |||
| 渡邊 賢也 | 1751129: M, 1回目発表 | 自然言語処理学 | 松本 裕治 |
|
title:Named Entity Recognition for Life science papers
abstract:Chemical papers are increasing too much to read by human. So it is important to make system to pick up relatesd papers. A general process for paper analysis in natural language processing is named entity recognition and relation extraction. In my presentation, I introduce performance of named entity recognition for life science papers. language of the presentation: Japanese | |||
| 青谷 拓海 | 1751001: M, 1回目発表 | 知能システム制御 | 杉本 謙二 |
|
title: Bottom-up Multi-agent Reinforcement Learning with Reward Sharing
abstract: Multi-agent reinforcement learning (MARL) is a frame work to make multiple agents (robots) in the same environment learn their policies simultaneously using reinforcement learning. In the conventional MARL, although decentralization is essential for feasible learning, rewards for the agents have been given from a centralized system (named as top-down MARL). To achieve the completely distributed autonomous systems, we tackle a new paradigm named bottom-up MARL, where the agents get respective rewards. Bottom-up MARL requires to share the respective rewards for creating orderly group behaviors, and therefore, methods to do so were investigated through simulations. We found that the orderly group behaviors could be created by considering the relationship between the agents. language of the presentation: Japanese | |||
| 井川 優太郎 | 1751004: M, 1回目発表 | 知能システム制御 | 杉本 謙二 |
|
title: Development of Biomechanical Energy Harvester Applicable to Various Movements
abstract: We developed a biomechanical energy harvester that can harvest electrical energy from human motions.Previously proposed devices are suitable only for limited movements (e.g., steady walking) due to the fixed reduction ratio from human movements to the power generator.On the other hand, our harvester equips a continuously variable transmission (CVT) system to be applicable for various movements. We present the concept of our harvester and a first prototype. language of the presentation: Japanese | |||
| 金子 拓光 | 1751031: M, 1回目発表 | 知能システム制御 | 杉本 謙二 |
|
title: A Data-driven Modeling Approach for Industrial Plants using Kalman Variational Auto-Encoders
abstract: An autonomous control algorithm for industrial plants is desired because a raising cost of operator is high and recovery work on the occurrence of abnormality is hard. Conventional methods were based on time series modeling techniques or anomaly detection techniques in order to detect abnormal condition sign. With the recent development of IoT and AI, the big industrial plant data is being collected and methods for model learning and control of dynamical systems from raw data are being proposed. Kalman Variational Auto-Encoders (KVAE) is one of the recent methods. This presentation introduces KVAE and show simulation results using KVAE. language of the presentation: Japanese | |||
| 佐々木 光 | 1751049: M, 1回目発表 | 知能システム制御 | 杉本 謙二 |
|
title: Gaussian Process Internal-State Policy Search
abstract: Internal state policy is available action determiner when occlusion occurs in observation. Internal state is obtained by extracting the information necesarry for action decision from past observation. In this study, we use Gaussian Process State-Space Model (GPSSM) as internal state policy. GPSSM makes the transition function nonlinear by assuming Gaussian process prior. We derive reinforcement learning algorithm using GPSSM that is optimized by EM policy search. In real world reinforcement learning problems, the perfect state information is rarely available due to sensor noise or occlusion. To cope with this problem, a control policy with internal state is beneficial since the internal state can be optimized as a memory of past observations as complementary to the imperfect state. In this study, we propose a novel internal-state policy model based on Gaussian Process State Space Models (GPSSMs) where we place a Guassian process prior on the space of internal-state transition functions. We show an RL algorithm with the policy model and simulation results. language of the presentation: Japanese | |||
| 杉野 峻生 | 1751056: M, 1回目発表 | 知能システム制御 | 杉本 謙二 |
|
title: Continuous Multi-task Learning by Structured Reservoir Computing
abstract: In recent years, neural networks (NN) have made excellent outcomes on the problems which are difficult to analytically solve due to complexity. NN has a critical problem, called catastrophic forgetting, which memories for tasks already learned would be broken when a new task is learned additionally. This problem interferes with continuous learning required for autonomous robots. It is supposed that catastrophic forgetting is caused by rewriting the whole network by backpropagation of NN. This study therefore propose a method to avoid catastrophic forgetting using a model that learns only the output layer of the network, called reservoir computing (RC). Instead of giving the network randomly like the conventional RC, the network structure and its weights are designed based on a fractal complex network. This network modularized memories for multiple tasks and avoided the catastrophic forgetting. language of the presentation: Japanese | |||
| 多川 純平 | 1751062: M, 1回目発表 | 知能システム制御 | 杉本 謙二 |
|
title: Convergence Rate of Average Consensus in activity-driven networks
abstract: we investigate asymptotic properties of a consensus protocol taking place in a class of temporal (i.e., time-varying) networks called the activity driven network. We first show that a standard methodology provides us with an estimate of the convergence rate toward the consensus, in terms of the eigenvalues of a matrix whose computational cost grows exponentially fast in the number of nodes in the network. To overcome this difficulty, we then derive alternative bounds involving the eigenvalues of a matrix that is easy to compute. Our analysis covers the regimes of 1) sparse networks and 2) fast-switching networks. We numerically confirm our theoretical results by numerical simulations. language of the presentation: Japanese | |||